arcprime-research-1 vs o3-deep-research: Three Ways to Judge an AI Patent Tool

Any AI claim chart and prior art search tool should be judged on at least 3 dimensions:

1. Did it find the evidence?

2. Is the reasoning sound?

3. Can you trust every citation?

We just upgraded the search-and-mapping engine behind our claim charts and prior art search. Before we released it, the new version (arcprime-research-1) had to beat the old one (OpenAI's o3-deep-research) on all three dimensions.

Two AI judges graded every output blind. arcprime-research-1 found 60% more relevant prior art references and proved 50% more claim elements with a verified citation. The judges preferred its reasoning 4 out of 5 times. Its arguments supported its conclusions 80% more often, it was 65% better at pinpointing the strength of a mapping, and it leaned on the doctrine of equivalents 80% less.

On citations, we asked o3-deep-research to verify every one. It verified MOST, but not all, of the time. Unfortunately, we couldn’t control when it decided to get lazy. arcprime-research-1 enforces verification in code, so every citation gets checked, every time.

How did we achieve these results? The biggest levers were using a cleaner, agent-friendly search index and smarter models. Agent-friendly search indexes provide the cleanest data as input to the AI. Less noise going into the agent means it spends its effort reasoning instead of wading through junk. And the newest frontier models made a real difference across the board. We also discovered a few search tricks that we probably should keep to ourselves. Can’t give away the shop :)

One thing that we thought would work but did not work: when a claim element was still weakly supported, we told the engine exactly what was missing and made it look beyond the manufacturer's website: teardowns, regulatory filings, research papers, trade press. It followed orders, but it didn't find more evidence. In fact, it worsened results for the claim as a whole. Maybe much of what's missing just isn't public.

In building arcprime-research-1, it's clear to me that there will not be one supreme AI model to do every task. Certain art is not prior art based on its publication date. Certain elements matter more than others. A "controller" can be implied by some equipment, but a "programmable logic controller" can't. We've encoded these rules and hundreds more into arcprime-research-1 to generate better claim charts and prior art searches. It's a specialist agent focused on a specialist task.

But ask them to plan the best Halloween party, arcprime-research-1 would be destroyed by o3-deep-research.